Advanced RVC Inference

A state-of-the-art web UI crafted to streamline rapid and effortless RVC inference — featuring a model downloader, voice splitter, batch inference, training pipeline, real-time conversion, and a full CLI.

Note

If you want to use old version switch to v1 branch.

---

Features

Inference

• Voice Inference — Single & batch conversion, TTS, pitch shifting, formant shifting, audio cleaning, Whisper transcription
• Real-Time Conversion — Live mic voice conversion with VAD and low-latency processing
• 30+ F0 Methods — rmvpe, crepe, fcpe, harvest, hybrid, and many more
• F0 Autotune — Automatic pitch correction with configurable strength
• Audio Cleaning — Built-in denoising for cleaner output

Audio Processing

• Audio Separation — Vocal/instrumental isolation (MDX-Net, Roformer, BS-Roformer), karaoke, reverb removal, denoising
• Auto Pretrained Download — Automatically downloads pretrained models from HuggingFace

Training Pipeline

• End-to-End Training — Dataset creation → preprocessing → feature extraction → training → model export
• 🔧 Auto Model Download — Predictor (RMVPE/FCPE) and embedder (HuBERT) models download automatically before training starts — no more "model not found" errors!
• 4 Vocoders — HiFi-GAN NSF (Default), BigVGAN, MRF-HiFi-GAN, RefineGAN
• 7 Optimizers — AdamW, RAdam, AnyPrecisionAdamW, AdaBelief, AdaBeliefV2, Ranger2020, Prodigy
• Enhanced Loss Functions — Multi-scale STFT loss, phase loss, envelope loss, KL-divergence loss (from Codename RVC Fork v4)
• Training Enhancements — LR warmup, KL annealing, gradient clip scheduling, decoder freezing (from Codename RVC Fork v4)
• Robust Data Loading — Safe numpy loading with NaN/Inf handling, corrupted file recovery, increased sequence length limits (900→1800)
• Advanced Options — Gradient accumulation, torch.compile(), 8-bit Adam, cosine annealing LR, overtraining detection
• Architecture Support — RVC and SVC (from Vietnamese-RVC)
• Embedder Mix — Layer-wise embedding mixing with configurable ratios (from Vietnamese-RVC)
• 🚀 3× Faster Training — --fast_train flag bundles TF32 matmul + cuDNN benchmark + torch.compile + expandable_segments allocator. Vocal-quality-safe (no loss/numerics changes).
• 🚀 bf16 Auto-Mode — --bf16_adamw flag (Applio-parity shortcut) forces AnyPrecisionAdamW + bf16 autocast. Recommended on Ampere+ GPUs (A100/H100/RTX 30xx+/40xx+).

🔒 Security Hardening

• Safe Deserialization — All torch.load() calls route through safe_torch_load (forces weights_only=True). Restricted pickle.Unpickler whitelist blocks every known RCE gadget.
• Path Traversal Guards — validate_path_within() wired into 20+ os.path.join sites in inference + training.
• Hardened Downloaders — All downloaders enforce: 8 GB size cap, extension whitelist, filename sanitization, timeout=300s.
• No Silent Failures — Bare except: clauses replaced with typed exceptions.

Platform & Integration

• CLI — Full command-line interface via rvc-cli
• ZLUDA Support — Full AMD GPU support via ZLUDA
• XPU Support — Intel GPU support via XPU backend
• Push to Hub — Upload trained models directly to HuggingFace Hub
• 44 Languages — Full UI translation support

---

Supported Vocoders

Vocoder	Description	Pitch Required
Default (HiFi-GAN NSF)	HiFi-GAN with Neural Sine Filter. Recommended for best compatibility.	Yes
BigVGAN	Snake activations with Anti-Aliasing. State-of-the-art audio quality.	Yes
MRF-HiFi-GAN	HiFi-GAN with Multi-Receptive Field fusion. Richer feature extraction.	Yes
RefineGAN	U-Net based vocoder with parallel residual blocks. High-fidelity spectral detail.	Yes

---

Supported Optimizers

Optimizer	Category	Rating	Best For
AdamW	PyTorch Built-in	⭐⭐⭐⭐⭐	General-purpose, most reliable (default)
RAdam	PyTorch Built-in	⭐⭐⭐⭐	Warmup-free training, short training runs
AnyPrecisionAdamW	Mixed-Precision	⭐⭐⭐⭐	Bfloat16 training, long runs with Kahan summation
AdaBelief	Belief-Based	⭐⭐⭐	Better conditioned adaptive learning rates
AdaBeliefV2	Belief-Based	⭐⭐⭐	Stable deep training with AMSGrad + InverseSqrt scheduler
Ranger2020 🆕	Advanced	⭐⭐⭐⭐⭐	RAdam + Lookahead + Gradient Centralization
Prodigy 🆕	D-Adaptation	⭐⭐⭐⭐⭐	Automatic LR tuning (lr=1.0 works!)

🆕 Newly added from Codename RVC Fork v4

---

Getting Started

1. Install

git clone https://github.com/ArkanDash/Advanced-RVC-Inference.git
cd Advanced-RVC-Inference
pip install -r requirements.txt

Or install from PyPI:

pip install git+https://github.com/ArkanDash/Advanced-RVC-Inference.git

GPU Support (CUDA)

pip install git+https://github.com/ArkanDash/Advanced-RVC-Inference.git
pip install onnxruntime-gpu

ZLUDA (AMD GPU)

ZLUDA allows CUDA applications to run on AMD GPUs. Just install PyTorch with ZLUDA support — Advanced RVC will auto-detect and configure itself.

# Follow the ZLUDA installation guide for your AMD GPU
# Then install Advanced RVC normally — ZLUDA is auto-detected
pip install git+https://github.com/ArkanDash/Advanced-RVC-Inference.git

2. Run

# Launch the web UI
rvc-gui

# Or via Python module
python -m arvc.app.gui

# With a public share link
python -m arvc.app.gui --share

The interface will be available at http://localhost:7860.

3. CLI Usage

# Voice conversion
rvc-cli infer -m model.pth -i input.wav -o output.wav

# Audio separation
rvc-cli uvr -i song.mp3

# Show all commands
rvc-cli --help

Fast Training

# ~3× faster training, vocal-quality-safe
rvc-cli train my_model --fast_train true --epochs 200 --batch_size 4

# Additional ~1.5–2× speedup on Ampere+ GPUs
rvc-cli train my_model --fast_train true --bf16_adamw true --epochs 200 --batch_size 8

4. Google Colab

Notebook	Description
	Full Web UI
	CLI only — lightweight headless mode

---

Terms of Use

The use of the converted voice for the following purposes is strictly prohibited:

• Criticizing or attacking individuals
• Advocating for or opposing specific political positions, religions, or ideologies
• Publicly displaying strongly stimulating expressions without proper zoning
• Selling of voice models and generated voice clips
• Impersonation of the original owner of the voice with malicious intentions
• Fraudulent purposes that lead to identity theft or fraudulent phone calls

---

Credits

This project builds upon the work of many open-source projects and contributors.

• RVC
• Vietnamese-RVC
• PolTrain
• Applio
• Codename RVC Fork 4 (Private)
• python-audio-separator
• whisper
• BigVGAN
• ZLUDA
• bitsandbytes

---

License

This project is licensed under the MIT License — see the LICENSE file for details.